Global Optimality Conditions for Optimization Problems

Transcription

1 The 7th International Symposium on Operations Research and Its Applications (ISORA 08) Lijiang, China, October 31 Novemver 3, 2008 Copyright 2008 ORSC & APORC, pp Global Optimality Conditions for Optimization Problems Zhiyou Wu 1,2, 1 School of Mathematics and Computer Science, Chongqing Normal University, Chongqing , China 2 Current address: Institute for Statistics and Mathematical Economic Theory, University of Karlsruhe, Karlsruhe, Germany. Abstract We establish new necessary and sufficient optimality conditions for optimization problems. In particular, we establish tractable optimality conditions for the problems of minimizing a weakly convex or concave function subject to standard constraints, such as box constraints, binary constraints, and simplex constraints. Our main theoretical tool for establishing these optimality conditions is abstract convexity. Keywords Global optimization; optimality conditions; abstract convexity. 1 Introduction We study optimality conditions for some classes of global minimization problems. References [1]-[3] and [6, 7, 12], etc. study global optimality conditions for the problems with quadratic objective function subject to either box constraints, binary constraints, quadratic constraints, or linear constraints. In the present paper, we follow the approach of references [2, 3, 9, 12, 13], abstract convexity is used in order to obtain new optimality conditions for global optimization problems. These optimality conditions are expressed in terms of abstract subdifferential (L-subdifferential) and abstract normal cone (L-normal cone). The present paper extends the existing optimality conditions to the necessary and sufficient optimality condition for a general global optimization problem: (P) min f (x), f : x U Rn R {+ },U dom f. This necessary and sufficient condition given in this paper is expressed in terms of (ε,l)- subdifferential and (ε,l)-normal set (see Theorem 1), which extends classical results expressed in terms of ε-subdifferential and ε-normal cone in the sense of convex analysis for concave minimization problems, which can be found in [1, 11] (see Corollary 3.1). Our analysis allows us to obtain some tractable optimality conditions for the classes of problems the objective function is a weakly convex or concave function subject to arbitrary constraint set. A function f is a weakly convex (or concave ) function if This research was partially supported by Alexander von Humboldt Foundation. Corresponding author. zhiyouwu@263.net.

2 378 The 7th International Symposium on Operations Research and Its Applications and only if it can be represented as the sum of a quadratic function and a convex (or concave) function. These classes of functions are broad enough ( see Section 4). So it always not easy to get some tractable global optimality conditions for weakly convex (or concave) problems. Reference [13] gives some sufficient conditions for some special classes of weakly convex programming problems. The present paper gives more stronger sufficient optimality conditions for general weakly convex problems, which extends the results given in [13]. The present paper also gives some necessary optimality conditions for general weakly concave problems (see Theorem 13), which are new results given in this paper. The lay-out of the paper is as follows. Section 2 provides a necessary and sufficient condition in terms of (ε, L)-subdifferential and (ε, L)-normal set. Some sufficient and/or necessary conditions in terms of L-subdifferential and L-normal cone are also presented in Section 2. Section 3 provides sufficient global optimality conditions for the class of problems the objective function a weakly convex function. Section 4 provides necessary optimality conditions for the class of problems the objective function is a weakly concave function. We use the following notation: R n is the n-dimensional Euclidean space; R n + is the n-dimensional non-negative Euclidean space, i.e., R n + := {(x 1,...,x n ) R n x i 0,i = 1,...,n}; u,v stands for the inner product of vectors u and v; S n is the set of symmetric n n matrices; Diag(q) is a diagonal matrix with the diagonal q; A B means that the matrix A B is positive semidefinite. f (x) and ε f (x) are the subdifferential and ε-subdifferential respectively of a convex function f at a point x in the sense of convex analysis. N U (x) and NU ε (x) are the normal cone and ε-normal set respectively of a convex set U at a point x in the sense of convex analysis. 2 Necessary and sufficient conditions for global optimality In this section we first give some preliminaries from abstract convexity (see [5, 8, 10]). Let X be a set and H be a set of functions h : X R. Let L be a set of functions defined on R n. For a function f : R n R {+ } and a point x dom f = {x R n : f (x) < + } define L-subdifferential L f ( x) of f at x by L f ( x) = {l L : f (x) f ( x) + l(x) l( x), x R n }. For ε 0, define (ε,l)-subdifferential ε,l f ( x) of f at x by ε,l f ( x) = {l L : f (x) f ( x) + l(x) l( x) ε, x R n }. Obviously, L f ( x) ε,l f ( x), for any ε 0.

3 Global Optimality Conditions for Optimization Problems 379 Let U R n and let x U. The L-normal set and the (ε,l) normal set of U at x are defined as follows: and N L,U ( x) = {l L : l(x) l( x), x U} N ε L,U( x) = {l L : l(x) l( x) + ε, x U}. If L is a cone, that is, (l L,λ > 0) = (λl L) then N L,U ( x) is also a cone. Since the set L we consider is always a subspace(in particular, a cone), we will use the usual term L- normal cone instead of L-normal set. Properties of (ε, L)-subdifferential (L-subdifferential) and (ε,l)-normal set (L-normal set) have been investigated in ([4, 8]). L-subdifferentials ((ε, L)-subdifferentials) and L-normal cones ((ε, L)-normal sets) are main tools for deriving global optimality conditions for a global minimizer. In other words L := {l : l = 12 } Qx,x + β,x with Q = Diag(q); q,β Rn. (1) We call L the subset of L which consists of functions l(x) = α x 2 + β,x with α R. We now give a necessary and sufficient global optimality condition with (ε,l)-subdifferential and (ε,l)-normal set for the following general global optimization problem (P), which extends the necessary and sufficient global optimality conditions given with ε-subdifferential of convex functions and ε-normal set of convex sets for concave minimization problems (see [1, 11]): (P) minimize f (x) subject to x U, f : R n R {+ } and U dom f = {x R n : f (x) < + }. Theorem 1. Let L be a subspace of real-valued functions defined on R n. Let x U. Suppose that f is H L -convex on U. Then x is a global minimizer of problem (P) if and only if ε,l ( f )( x) N ε L,U( x), ε > 0, (2) H L = {l(x) + c l L,c R}. (3) Here we give a necessary and sufficient global optimality condition for the general optimization problem (P) in terms of (ε, L)-subdifferentials and (ε, L)-normal set. If U is convex and f (x) is concave on U, then the following results which can be found in [1, 11] can be reduced from Theorem 1. Corollary 2. Let x U. Suppose that U is a convex set and f is a concave function on U. Then x is a global minimizer of problem (P) if and only if ε ( f )( x) N ε U( x), ε > 0. (4)

4 380 The 7th International Symposium on Operations Research and Its Applications Remark 1. From Theorem 1, we can also obtain the following results: Let L be a subspace of real-valued functions defined on R n. Let x U. (i). If x is a global minimizer of problem (P), then ε,l ( f )( x) N ε L,U( x), ε 0; (5) (ii). If for any x U, L ( f )(x) /0, then x is a global minimizer of problem (P) if and only if (5) holds. Indeed, from the first part of the proof of Theorem 1 (for the necessary condition), we know that if x is a global minimizer of problem (P), then (5) holds, f does not require to be H L -convex on U. Since L ( f )(x) /0, x U implies that f is H L -convex on U, (ii) can be obtained directly from Theorem 1. In general, it is very difficult to calculate the (ε, L)-subdifferential and the (ε, L)- normal set for general function f and general set U, and even if f and U are convex, it is also very difficult to calculate the ε-subdifferential and the ε-normal set in the sense of convex analysis. But, for some kinds of functions f and for some kinds of sets U, we can calculate the L-subdifferential and the L-normal cone. The following simple results will be useful in the sequel. Corollary 3. (Necessary Condition) Let L be a subspace of real-valued functions defined on R n. Let x U. If x is a global minimizer of problem (P), then L ( f )( x) N L,U ( x). (6) The following sufficient condition has been presented in [2, 3], which study sufficient conditions for quadratic functions over binary or/and box sets. Proposition 4. (Sufficient Condition) Let L be a subspace of real-valued functions defined on R n. Let x U. If ( L f ( x) ) N L,U ( x) /0 (7) then x is a global minimizer of problem (P). Corollary 3 and Proposition 4 imply that ( L f ( x) ) N L,U ( x) /0 = ε,l ( f )( x) N ε L,U( x), ε > 0. Generally, there is a gap between the sufficient condition (7) and the necessary condition (6). But in some cases, both of (7) and (6) are necessary and sufficient conditions. In Section 7, we will see that if f is a quadratic function, then the sufficient condition (7) and the necessary condition (6) are equivalent in some sense. Proposition 5. Let L be a subspace of real-valued functions defined on R n, x U. If f L, then ( ) L f ( x) N L,U ( x) /0 f N L,U ( x) L ( f )( x) N L,U ( x). (8) Condition (8) means that if f L, both of (7) and (6) are necessary and sufficient conditions.

5 Global Optimality Conditions for Optimization Problems Sufficient Conditions for Weakly Convex Problems Consider problem U dom f, (P 1 ) minimize f (x) subject to x U, with A 0 S n and p is a proper convex function. f (x) = 1 2 A 0x,x + p(x), (9) Proposition 6. Let f be a function defined by (9). Let x dom p, the domain of p. Then L f ( x) {l L l(x) = 1 2 Qx,x + β,x : A 0 Q 0,(Q A 0 ) x + β p( x)}. (10) Corollary 7. Let f (x) = 1 2 A 0x,x + a 0,x be a quadratic function. Then L f ( x) = {l L l(x) = 1 2 Qx,x + β,x : A 0 Q 0,β = a 0 + (A 0 Q) x}.(11) Theorem 8. Let U R n and x U. If [SC1] q, b R n such that A 0 Diag(q), b A 0 x p( x) and l q,b, x N L,U( x). Then x is a global minimizer of problem (P 1 ). Corollary 9. Let f be a quadratic function, f (x) = 1 2 A 0x,x + a 0,x. If [SCQ1] q R n such that A 0 Diag(q), b = a 0 + A 0 x and l q,b, x N L,U( x), then x is a global minimizer of problem (P 1 ). Furthermore, ( ) [SCQ1] L f ( x) N L,U( x) /0. We can obtain tractable sufficient conditions if the normal cone N L,U( x) can be easily calculated, in particular, if U is one of the following sets, then we have the following results: Corollary 10. Let U = n i=1 {u i,v i } and x U. If there exist q,b R n such that A 0 Diag(q) and b A 0 x p( x) and Xb 1 Diag(q)(v u), 2 x i : = 0 if x i = u i = v i 1 if x i = u i 1 if x i = v i X : = Diag( x 1,..., x n ) (12) and v = (v 1,...,v n ),u = (u 1,...,u n ), then x is a global minimizer of problem (P 1 ).

6 382 The 7th International Symposium on Operations Research and Its Applications Corollary 11. Let U = {x = (x 1,...,x n ) : u i x i v i } and x U. If there exist q,b R n such that A 0 Diag(q) and b A 0 x p( x) and X b b 1 2 Q (v u)}, x i (b) : = 0 if x i = u i = v i 1 if x i = u i < v i 1 if x i = v i > u i b i if x i (u i,v i ) X b : = Diag( x 1 (b),..., x n (b)) (13) and v = (v 1,...,v n ),u = (u 1,...,u n ), then x is a global minimizer of problem (P 1 ). Corollary 12. Let U = co{v 1,...,v k },k n + 1 be a simplex in R n and x U. If there exist q,b R n such that A 0 Diag(q), b A 0 x p( x) and then x is a global minimizer of problem (P). max i=1,...,k h q,b (v i) min x U h+ q,b (x), (14) 4 Necessary Conditions for Weakly Concave Problems Consider the problem (P 2 ) minimize f (x) subject to x U, U dom f, f (x) = 1 2 A 0x,x + p(x), with A 0 S n and p is a proper concave function. A function f is called to be a weakly concave function if f is a weakly convex function. Theorem 13. Let U R n, x U. If x is a global minimizer of problem (P 2 ), then [NC1] l q,b, x N L,U( x) for any q,b R n with Diag(q) A 0,A 0 x b ( p)( x), Q = Diag(q). l q,b, x (x) := 1 2 n i=1 q i x 2 i + n i=1 b i x i n i=1 q i x i x i = 1 Qx,x + b Q x,x, (15) 2 We can derive some necessary conditions for the following optimization problems (QP) with quadratic objective functions: (QP) minimize f (x) subject to x U, (16) f (x) := 1 2 A 0x,x + a 0,x, and U R n.

7 Global Optimality Conditions for Optimization Problems 383 Corollary 14. Let U R n, x U. If x is a global minimizer of problem (QP), then [NCQ1] l q,b, x N L,U( x) for any q R n with Diag(q) A 0,b = a 0 + A 0 x. Furthermore, [NCQ1] L( f )( x) N L,U( x). Remark 2. For quadratic problem (QP), conditions [SCQ1],[NCQ1] and [NCQ2] have the following relationships: [SCQ1] = [NCQ1], and [NCQ1] from Remark 7.1 = [SCQ1] = [NCQ2], but [NCQ2] [SCQ1] [NCQ1] from Remark 7.1 = If U = U( x) = n i=1 U i( x) for x U, then Acknowledgments [SCQ1] [NCQ2], but [NCQ2] [NCQ1]. [NCQ2] = [NCQ1]. The authors are very thankful to Professor A. M. Rubinov who made some helpful suggestions for this paper. References [1] J. B. Hiriart-Urruty and C. Lemarechal, Testing necessary and sufficient conditions for global optimality in the problem of maximizing a convex quadratic over a convex polyhedron, Preliminary report, Seminar of Numerical Analysis, University Paul Sabatier, Toulouse, [2] V. Jeyakumar, A. M. Rubinov, Z.Y. Wu. Sufficient global optimality conditions for non-convex quadratic minimization problems with box constraints, Journal of Global Optimization, 36(3), , [3] V. Jeyakumar, A. M. Rubinov, Z.Y. Wu. Global Optimality Conditions for Non-convex Quadratic Minimization Problems with Quadratic Constraints, Mathematical Programming, Series A, 110, , [4] V. Jeyakumar, A. M. Rubinov, Z.Y. Wu. Generalized Fenchel s Conjugation Formula and Duality for Abstract Convex Functions, Journal of Optimization Theory and Applications, 132, , [5] D. Pallaschke and S. Rolewicz, Foundations of Mathematical Optimization, Kluwer Academic Publishers, Dordrecht, [6] J.M. Peng and Y. Yuan, Optimization conditions for the minimization of a quadratic with two quadratic constraints, SIAM J. Optim. 7(3) (1997), [7] M. C. Pinar, Sufficient global optimality conditions for bivalent quadratic optimization, J. Optim. Theor. Appl., 122(2) (2004), [8] A. M Rubinov, Abstract Convexity and Global Optimization, Kluwer Academic Publishers, 2000.

8 384 The 7th International Symposium on Operations Research and Its Applications [9] A.M. Rubinov, Z. Y. Wu, Optimality conditions in global optimization and their applications, Mathematical Programming, Series B, Published online: [10] I. Singer, Abstract convex analysis, John Wiley & Sons, Inc., New York, [11] H. Tuy, Convex Analysis and Global Optimization, Kluwer Academic publishers, Dordrecht/Boston/London, [12] Z.Y. Wu, V. Jeyakumar, A. M. Rubinov, Sufficient Conditions for Globally Optimality of Bivalent Nonconvex Quadratic Programs, Journal of Optimization Theory and Applications,133, , [13] Z.Y. Wu, Sufficient Global Optimality Conditions for Weakly Convex Minimization Problems, Journal of Global Optimization, 39(3), , 2007.